Unverified50% confidenceOpinionExact time
The developer proposes a hypothesis: models defend familiar facts and abandon surprising ones.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedKnowledge Kernel采用Reality→Evidence→Facts→Reasoning的自上而下分层模型，将推理明确排除在事实层之外73% similarUnverified将人类已确认意图与模型原始推断分离后，可以判断错误出在「人没说清」还是「模型没理解」71% similarUnverified当研究者尝试给Video-LLM注入物理知识时，错误合理化反而加剧，模型变得更擅长编造理由解释违规现象70% similarUnverifiedThe theoretical foundation of this method comes from the counterfactual intervention framework in causal reasoning.70% similarUnverifiedBy injecting fabricated reasoning, it is possible to forge a model's thought process and induce it to output content it otherwise should not, constituting a form of jailbreak attack.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1019904API
curl https://kongchang.com/api/v1/knowledge/claims/1019904MCP
get_claim(id=1019904)